LJ Speech Tacotron-1 Models trained with latest commits

TTS (Text-to-Speech)
#1

Hello,
I have trained a Tacotron1 model with MozillaTTS project and I have got some good results. In order to compare my results with the models trained by other people to make sure that the model is trained as just expected, I have searched for a published models page. When I had a look at released models page in Mozilla TTS: https://github.com/mozilla/TTS/wiki/Released-Models

I have found out that these models were trained 2 years ago and I want to make sure if is there any other page for newer models published for LJSpeech database.

Best Regards.

#2

Do you have Tensorboard logs?

#3

No they are all we released so far. But we can also add your model to that list if you like?

#4

yes, i have training logs

#5

Thank you, After making sure that the models generate good quality synthesis for any kind of text inputs, I can share the model trained with the latest project commits

#6

Please find the logs attached.
logs_ljspeech.zip (889.6 KB)

#7

I think it might be more telling if you can point Tensorboard to the dir the model is in, so the can read the event file and then you can get a picture of how the training progressed. For reference, this is a healthy training https://github.com/mozilla/TTS/issues/337

In general, if inference is good and you get good speech, you should be good. How large is the dataset and of what quality?

#8

I directly used LJSpeech Database. The general quality of the voice is good. The issue is that I need to compare it with a prebuilt model performance to make sure I’am doing everything correctly